Hostile behavior detection from multiple view points using RGB-D sensor
Amol Patwardhan · 2017
This paper presents a novel method for the detection of hostile behavior by a person from multiple viewing directions. Three dimensional, kinematic and gesture frequency based features were extracted from video and depth input data streams. Silhouette based features were extracted to compensate for instances where sensor tracked data was unavailable. The features were used to train the supervised learning classifier using 10-fold cross validation. The hostile behavior detection method was evaluated using supervised classification techniques such as support vector machine, random forest, Naïve Bayes and Multilayer Perceptron. The method showed overall accuracy of 90.273% using Random Forest classification. The method was compared to existing state of the art and showed better precision for specific viewing directions (full frontal, left and right profile view).